Relaxed Structure Tensor Representation for Robust Ori- ented Object Detection

TMLR Paper5970 Authors

Abstract:

Oriented object detection predicts oriented bounding boxes. Precisely predicting their orientation remains challenging due to angular periodicity, which introduces boundary discontinuity issues and symmetry ambiguities.
In this paper, we introduce Relaxed Structure Tensor Bounding Boxes (RST-BB), a representation inspired by classical image structure tensors encoding object orientation in addition to height and width.
RST-BB provides a simple yet efficient angle-coder approach that is robust to angular issues, effectively addresses square objects, and requires no additional hyperparameters. Extensive evaluations across five datasets demonstrate that RST-BB achieves state-of-the-art results with high angular prediction precision, establishing relaxed structure tensors as a robust and modular alternative for encoding orientation in oriented object detection. We make our code publicly available for seamless integration into existing detectors.

Submission Length: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Stephen_Lin1

Submission Number: 5970